Language models (LMs) trained on vast quantities of text data can acquire sophisticated skills such as generating summaries, answering questions or generating code. However, they also manifest behaviors that violate human preferences, e.g., they can generate offensive content, falsehoods or perpetuate social biases. In this thesis, I explore several approaches to aligning LMs with human preferences. First, I argue that aligning LMs can be seen as Bayesian inference: conditioning a prior (base, pretrained LM) on evidence about human preferences (Chapter 2). Conditioning on human preferences can be implemented in numerous ways. In Chapter 3, I investigate the relation between two approaches to finetuning pretrained LMs using feedback given by a scoring function: reinforcement learning from human feedback (RLHF) and distribution matching. I show that RLHF can be seen as a special case of distribution matching but distributional matching is strictly more general. In chapter 4, I show how to extend the distribution matching to conditional language models. Finally, in chapter 5 I explore a different root: conditioning an LM on human preferences already during pretraining. I show that involving human feedback from the very start tends to be more effective than using it only during supervised finetuning. Overall, these results highlight the room for alignment techniques different from and complementary to RLHF.

Multi-modal relation extraction (MMRE) is a challenging task that aims to identify relations between entities in text leveraging image information. Existing methods are limited by their neglect of the multiple entity pairs in one sentence sharing very similar contextual information (ie, the same text and image), resulting in increased difficulty in the MMRE task. To address this limitation, we propose the Variational Multi-Modal Hypergraph Attention Network (VM-HAN) for multi-modal relation extraction. Specifically, we first construct a multi-modal hypergraph for each sentence with the corresponding image, to establish different high-order intra-/inter-modal correlations for different entity pairs in each sentence. We further design the Variational Hypergraph Attention Networks (V-HAN) to obtain representational diversity among different entity pairs using Gaussian distribution and learn a better hypergraph structure via variational attention. VM-HAN achieves state-of-the-art performance on the multi-modal relation extraction task, outperforming existing methods in terms of accuracy and efficiency.

Additionally, it explores how LLMs can be more closely aligned with human preferences through innovative reinforcement learning frameworks and other novel methods that incorporate human feedback. The article also examines the emerging technique of retrieval augmented generation, integrating external knowledge into LLMs. The ethical dimensions of LLM deployment are discussed, underscoring the need for mindful and responsible application. Concluding with a perspective on future research trajectories, this review offers a succinct yet comprehensive overview of the current state and emerging trends in the evolving landscape of LLMs, serving as an insightful guide for both researchers and practitioners in artificial intelligence.

This paper introduces Interactive Tables (iTBLS), a dataset of interactive conversations situated in tables from scientific articles. This dataset is designed to facilitate human-AI collaborative problem-solving through AI-powered multi-task tabular capabilities. In contrast to prior work that models interactions as factoid QA or procedure synthesis, iTBLS broadens the scope of interactions to include mathematical reasoning, natural language manipulation, and expansion of existing tables from natural language conversation by delineating interactions into one of three tasks: interpretation, modification, or generation. Additionally, the paper presents a suite of baseline approaches to iTBLS, utilizing zero-shot prompting and parameter-efficient fine-tuning for different computing situations. We also introduce a novel multi-step approach and show how it can be leveraged in conjunction with parameter-efficient fine-tuning to achieve the state-of-the-art on iTBLS; outperforming standard parameter-efficient fine-tuning by up to 15% on interpretation, 18% on modification, and 38% on generation.

In Simultaneous Machine Translation (SiMT) systems, training with a simultaneous interpretation (SI) corpus is an effective method for achieving high-quality yet low-latency systems. However, it is very challenging to curate such a corpus due to limitations in the abilities of annotators, and hence, existing SI corpora are limited. Therefore, we propose a method to convert existing speech translation corpora into interpretation-style data, maintaining the original word order and preserving the entire source content using Large Language Models (LLM-SI-Corpus). We demonstrate that fine-tuning SiMT models in text-to-text and speech-to-text settings with the LLM-SI-Corpus reduces latencies while maintaining the same level of quality as the models trained with offline datasets. The LLM-SI-Corpus is available at \url{https://github.com/yusuke1997/LLM-SI-Corpus}.

We consider the problem of estimating probability density functions based on sample data, using a finite mixture of densities from some component class. To this end, we introduce the $h$-lifted Kullback--Leibler (KL) divergence as a generalization of the standard KL divergence and a criterion for conducting risk minimization. Under a compact support assumption, we prove an $\mc{O}(1/{\sqrt{n}})$ bound on the expected estimation error when using the $h$-lifted KL divergence, which extends the results of Rakhlin et al. (2005, ESAIM: Probability and Statistics, Vol. 9) and Li and Barron (1999, Advances in Neural Information ProcessingSystems, Vol. 12) to permit the risk bounding of density functions that are not strictly positive. We develop a procedure for the computation of the corresponding maximum $h$-lifted likelihood estimators ($h$-MLLEs) using the Majorization-Maximization framework and provide experimental results in support of our theoretical bounds.

Critical Infrastructure Facilities (CIFs), such as healthcare and transportation facilities, are vital for the functioning of a community, especially during large-scale emergencies. In this paper, we explore a potential application of Large Language Models (LLMs) to monitor the status of CIFs affected by natural disasters through information disseminated in social media networks. To this end, we analyze social media data from two disaster events in two different countries to identify reported impacts to CIFs as well as their impact severity and operational status. We employ state-of-the-art open-source LLMs to perform computational tasks including retrieval, classification, and inference, all in a zero-shot setting. Through extensive experimentation, we report the results of these tasks using standard evaluation metrics and reveal insights into the strengths and weaknesses of LLMs. We note that although LLMs perform well in classification tasks, they encounter challenges with inference tasks, especially when the context/prompt is complex and lengthy. Additionally, we outline various potential directions for future exploration that can be beneficial during the initial adoption phase of LLMs for disaster response tasks.

Accurate spatio-temporal information about the current situation is crucial for smart city applications such as modern routing algorithms. Often, this information describes the state of stationary resources, e.g. the availability of parking bays, charging stations or the amount of people waiting for a vehicle to pick them up near a given location. To exploit this kind of information, predicting future states of the monitored resources is often mandatory because a resource might change its state within the time until it is needed. To train an accurate predictive model, it is often not possible to obtain a continuous time series on the state of the resource. For example, the information might be collected from traveling agents visiting the resource with an irregular frequency. Thus, it is necessary to develop methods which work on sparse observations for training and prediction. In this paper, we propose time-inhomogeneous discrete Markov models to allow accurate prediction even when the frequency of observation is very rare. Our new model is able to blend recent observations with historic data and also provide useful probabilistic estimates for future states. Since resources availability in a city is typically time-dependent, our Markov model is time-inhomogeneous and cyclic within a predefined time interval. To train our model, we propose a modified Baum-Welch algorithm. Evaluations on real-world datasets of parking bay availability show that our new method indeed yields good results compared to methods being trained on complete data and non-cyclic variants.

This paper introduces a novel approach, evolutionary multi-objective optimisation for fairness-aware self-adjusting memory classifiers, designed to enhance fairness in machine learning algorithms applied to data stream classification. With the growing concern over discrimination in algorithmic decision-making, particularly in dynamic data stream environments, there is a need for methods that ensure fair treatment of individuals across sensitive attributes like race or gender. The proposed approach addresses this challenge by integrating the strengths of the self-adjusting memory K-Nearest-Neighbour algorithm with evolutionary multi-objective optimisation. This combination allows the new approach to efficiently manage concept drift in streaming data and leverage the flexibility of evolutionary multi-objective optimisation to maximise accuracy and minimise discrimination simultaneously. We demonstrate the effectiveness of the proposed approach through extensive experiments on various datasets, comparing its performance against several baseline methods in terms of accuracy and fairness metrics. Our results show that the proposed approach maintains competitive accuracy and significantly reduces discrimination, highlighting its potential as a robust solution for fairness-aware data stream classification. Further analyses also confirm the effectiveness of the strategies to trigger evolutionary multi-objective optimisation and adapt classifiers in the proposed approach.

The Socratic method is a way of guiding students toward solving a problem independently without directly revealing the solution to the problem. Although this method has been shown to significantly improve student learning outcomes, it remains a complex labor-intensive task for instructors. Large language models (LLMs) can be used to augment human effort by automatically generating Socratic questions for students. However, existing methods that involve prompting these LLMs sometimes produce invalid outputs, e.g., those that directly reveal the solution to the problem or provide irrelevant or premature questions. To alleviate this problem, inspired by reinforcement learning with AI feedback (RLAIF), we first propose a data augmentation method to enrich existing Socratic questioning datasets with questions that are invalid in specific ways. Next, we propose a method to optimize open-source LLMs such as LLama 2 to prefer ground-truth questions over generated invalid ones, using direct preference optimization (DPO). Our experiments on a Socratic questions dataset for student code debugging show that a DPO-optimized 7B LLama 2 model can effectively avoid generating invalid questions, and as a result, outperforms existing state-of-the-art prompting methods.